[WCCF] AMD Radeon R9 390X Pictured

[RussianSensation]

By moving HBM1 closer to the GPU die and reducing the PCB footprint, AMD could create a card 50-60% the size of an R9 295X2. Cover the HBM1 memory and the GPU die with a heatspreader (Let's call this the entire "package"). Then they can drop a small water-block on top of the package and the VRMs/Mosfets and use a 120mm AIO CLC to cool the entire thing.

The side benefit to this is AMD could find yet another 20-25W of power savings since there is a relationship between an ASIC's power usage and its operating temperatures, per IDC's image.

If HBM1 + reduced memory controller on the GPU die together save 30-40W power, AIO CLC save another 20-25W of power, and high density stack saves 20% of power over 290X's design (i.e., 50-60W), we are looking at 100-125W of power saved against a reference R9 290X.

HDS allows 20% power savings at the same frequency.

That means realistically speaking if Fiji XT is a 290-300W TDP card, AMD can use 100-125W of "extra power" headroom for performance/larger die. At that point AMD will need to figure out a way how to convert 100-125W of newly found power reserve to make a card 40-50% faster than the R9 290X.

 

[3DVagabond]

I am the latter (to hell with analog circuit design!) I've made it clear I don't work on video drivers, but I've been around long enough to have some knowledge. Hacks are always bad. The way drivers currently optimize for individual games....should be avoided. It's terrible design. I've seen the same kind of nonsense done for DVDs, HD DVDs and BluRays to fix stamping screw ups. It's never a reliable fix. It's never that easy to maintain.

I don't get the impression that they are talking about per game optimizations. I think they are referring to something more global.

 

[EndlessWargasm]

It seems to me that, if it is possible at all, that it would make sense to design the interposer to support 8192 bits of width for a number of reasons.

Look out, lots of assumptions from someone with no real technical knowledge.

I assume it wouldn't be any more expensive to manufacture and that there is more than enough room for the routing in a ~60x60mm interposer. And therefore there would be room harvest dies.

I'm also assuming that there is going to be some limits on the number of interposers that they can manufacture, especially to begin with. So being able to take failed 8192 bit 8 stack parts and use them as 6144 bit 6 stack parts say or 4096 or 2048 makes sense.

Over the next 12-18 months till 14/16nm become available in volume, AMD is going to going to need a stack of interposers and full stack of new GPUs, firepros and mobile parts especially.

But as I said, I don't know anything.

 

[flash-gordon]

If this engineering put into frame buffer utilization becomes real, then we must see some performance boost for 290(X) and specially 280(X)/285.

 

[LTC8K6]

It seems to me that, if it is possible at all, that it would make sense to design the interposer to support 8192 bits of width for a number of reasons.

Look out, lots of assumptions from someone with no real technical knowledge.

I assume it wouldn't be any more expensive to manufacture and that there is more than enough room for the routing in a ~60x60mm interposer. And therefore there would be room harvest dies.

I'm also assuming that there is going to be some limits on the number of interposers that they can manufacture, especially to begin with. So being able to take failed 8192 bit 8 stack parts and use them as 6144 bit 6 stack parts say or 4096 or 2048 makes sense.

Over the next 12-18 months till 14/16nm become available in volume, AMD is going to going to need a stack of interposers and full stack of new GPUs, firepros and mobile parts especially.

But as I said, I don't know anything.

IIRC, AMD said the interposer was made on an older established fab process, and did not require the latest technology to manufacture.

This may be why the first gen will only have 4GB of vram.

 

[Headfoot]

If this engineering put into frame buffer utilization becomes real, then we must see some performance boost for 290(X) and specially 280(X)/285.

Unless the solution involves some sort of fixed function hardware related to HBM or otherwise on the new one. It would be nice to see some free framebuffer capacity "increase" for the older sub 4gb cards. Especially the 285.

 

[Headfoot]

i think the problem is the routing
1024 pins per stack
1024 * 4 = 4096 traces

now double that and you get

8192 traces, that have to be routed in a 70mm x 70mm square

and probably also doubling power consumption

That's why its on an interposer, not PCB. Interposer is essentially routing layer of an actual die. So you can write vastly more routes at vastly smaller sizes vs PCB, even with the interposer being built on 65nm.

So what you're saying isnt true for interposer

 

[tviceman]

Thanks so much for this graph. I always thought the correlation between temps and power consumption was pretty low, but this confirms it for me. Looks like a little less than 1 watt per 3 degrees Celsius. Need good water cooling to bring about a significant (15+ watts) power savings at same clock speeds and voltage.

 

[DiogoDX]

By moving HBM1 closer to the GPU die and reducing the PCB footprint, AMD could create a card 50-60% the size of an R9 295X2. Cover the HBM1 memory and the GPU die with a heatspreader (Let's call this the entire "package"). Then they can drop a small water-block on top of the package and the VRMs/Mosfets and use a 120mm AIO CLC to cool the entire thing.

The side benefit to this is AMD could find yet another 20-25W of power savings since there is a relationship between an ASIC's power usage and its operating temperatures, per IDC's image.

If HBM1 + reduced memory controller on the GPU die together save 30-40W power, AIO CLC save another 20-25W of power, and high density stack saves 20% of power over 290X's design (i.e., 50-60W), we are looking at 100-125W of power saved against a reference R9 290X.

HDS allows 20% power savings at the same frequency.

That means realistically speaking if Fiji XT is a 290-300W TDP card, AMD can use 100-125W of "extra power" headroom for performance/larger die. At that point AMD will need to figure out a way how to convert 100-125W of newly found power reserve to make a card 40-50% faster than the R9 290X.

The problem is that we don't know if AMD will use carizo power save features in Fiji.

They have a power roadmap and I believe that the features "inter-frame power gating", "per part adaptative voltage" and "inteligent boost" must be similar to what Nvidia is doing in Maxwell to increase per/watt.

 

[DiogoDX]

Thanks so much for this graph. I always thought the correlation between temps and power consumption was pretty low, but this confirms it for me. Looks like a little less than 1 watt per 3 degrees Celsius. Need good water cooling to bring about a significant (15+ watts) power savings at same clock speeds and voltage.

I believe this varies with the different ASICs.

 

[ocre]

did you do the math. It is interesting.
Its not too much different on the low end.

but as it gets warmer we have
every 5 degrees is 4watts in the middle and the higher temps go to up 5 watts per 5 degree

So from the middle thats .8 watts per degree to 1watt per degree.

This is on 40nm. I would consider this best cast for 28nm. So keeping the card 20c cooler could actually shave off 15-20watts. That is just a rough estimate. 20nm may be more forgiving to temps though

 

[gamervivek]

Tom's found significant difference with LCS cards.

Also cryovenom,

http://www.tomshardware.com/reviews/cryovenom-r9-290-water-cooling-review,3756-11.html

 

[Justinbaileyman]

When are these new 3xx cards coming? Soon I hope? Anyone got the release date please?

 

[peateargryphon]

When are these new 3xx cards coming? Soon I hope? Anyone got the release date please?

Just rumors, but Computex (starting June 3rd) should reveal some official details. And most rumors point to all the new GPU releases happening in late June.

 

[Justinbaileyman]

Ohhh awesomeness...!! Thank you. Thats just right around the corner and this time I have the cash ready so I am so in for a new graphics card. I am really hoping that they are not all going to be water cooled like that picture on the first post of the r9 390x.

 

[Face2Face]

Thanks so much for this graph. I always thought the correlation between temps and power consumption was pretty low, but this confirms it for me. Looks like a little less than 1 watt per 3 degrees Celsius. Need good water cooling to bring about a significant (15+ watts) power savings at same clock speeds and voltage.

For more of a direct comparison, take a look at the NZXT G10 review @ Legit Reviews. Also keep in mind if the GPUs VRM was water cooled efficiency of of the MOSFETs would improve, resulting in less power consumption.

http://www.legitreviews.com/nzxt-kr...oler-review-on-an-amd-radeon-r9-290x_130344/4

The largest shock when testing the AMD Radeon R9 290X was the VDDC power output reading of the card in Watts. This is basically the calculation of GPU-Z 0.7.4 that shows the power output of the entire card in Watts. This also can be understood as how much power the card is using. Legit Reviews spoke with W1zzard over at TPU about this data
image: http://images.intellitxt.com/ast/adTypes/icon1.png
value and he said this is a reading provided directly by the voltage controller and that it likely is the result of multiplying the measured voltage with the measured current of the card. The Sapphire Radeon R9 290X with the stock cooler averaged 146 Watts with a peak of 171.5 Watts. The same card with the water cooler averaged just 120 Watts with a peak of 147 Watts. The fact that GPU-Z was using 26W less power with water cooling was astounding.We showed these results to AMD and they said that at lower temperatures that there will be less leakage across insulators inside the GPU. When talking about a temperature drop of 94C to 84C that difference is usually negligible, but since we had a 50C temperature drop that could be part of the reason we are seeing huge power savings.
Read more at http://www.legitreviews.com/nzxt-kr...d-radeon-r9-290x_130344/4#jrv5he6Zl92VUfBI.99

 

[sushiwarrior]

Hey, totally forgot about this forum haha. Just gonna clarify on the "big die" info I had posted earlier, which I didn't really state very clearly, I know it was eons ago. Had reliable knowledge of a >2500mm^2 die, with the extra tidbit that it was due to on-die memory. That's why I said the die size was way bigger than 550, but I should have said instead that it was package size or something. And on-die memory was what I meant when I said it wasn't more shaders. I was trying my hardest to be as vague as possible which didn't help, and I didn't have very detailed info myself. Believe me or not ()

 

[DiogoDX]

Hey, totally forgot about this forum haha. Just gonna clarify on the "big die" info I had posted earlier, which I didn't really state very clearly, I know it was eons ago. Had reliable knowledge of a >2500mm^2 die, with the extra tidbit that it was due to on-die memory. That's why I said the die size was way bigger than 550, but I should have said instead that it was package size or something. And on-die memory was what I meant when I said it wasn't more shaders. I was trying my hardest to be as vague as possible which didn't help, and I didn't have very detailed info myself. Believe me or not ()

You have some information about memory size and performance?

 

[gamervivek]

Hey, totally forgot about this forum haha. Just gonna clarify on the "big die" info I had posted earlier, which I didn't really state very clearly, I know it was eons ago. Had reliable knowledge of a >2500mm^2 die, with the extra tidbit that it was due to on-die memory. That's why I said the die size was way bigger than 550, but I should have said instead that it was package size or something. And on-die memory was what I meant when I said it wasn't more shaders. I was trying my hardest to be as vague as possible which didn't help, and I didn't have very detailed info myself. Believe me or not ()

You should be thankful to me for spreading it far and wide. :whiste:

Anyway, it took quite a while for it to materialize.

 

[dacostafilipe]

https://twitter.com/repi/status/601739457960763392

 

[Udgnim]

if AMD is sticking with blower style cooler and they get slammed for temperatures again, they only have themselves to blame

unless that's the WCE version, then the board looks really short

 

[Silverforce11]

So the leaked render is real.

Its tiny. Hopefully incredibly beastly.

Oh and it dumps all the waste heat out the case too? Take my money*!!

*If benches show great perf at 4K

 

[Face2Face]

Looks like the RADEON logo is going to glow red as well, similar to NVIDIA's reference cooler. I gotta say, it's a beauty and looks like a real premium product. Looking forward to seeing some benchmarks.

 

[monstercameron]

So the leaked render is real.

Its tiny. Hopefully incredibly beastly.

Oh and it dumps all the waste heat out the case too? Take my money*!!

*If benches show great perf at 4K

$7-8 Hundo is too rich for my blood, best I can do is $4-500 I wonder if there will be a cut down Fiji?

 

[Silverforce11]

$7-8 Hundo is too rich for my blood, best I can do is $4-500 I wonder if there will be a cut down Fiji?

Ofc there will be cut down Fiji.. what do you expect them to do with all the aborted dies, its AMD's first ever massive GPU die.